Step 3 of 9 · planned
Keep the hard cases
Four dogs isn't a test set. Build a dataset of deliberately awkward days that every future prompt has to survive.
What you'll build
A Braintrust dataset: the four dogs plus nasty ones — nothing logged, contradictory entries, a day that's only an incident.
Braintrust features
What you'll be able to show
“The awkward days are permanent now. Nothing can quietly break them again.”
🐾 Not written yet. Each step is written as the course is actually worked through, so it describes real scores, real regressions and real false alarms rather than a script guessed in advance.